Agent Harness Wiki

5 min read

Nightly Src Projects Desk Raw Survey (2026-05-02)

This raw note preserves the public-facing basis for the 2026-05-02 nightly src/ projects desk. It intentionally keeps the evidence layer compact and safe to publish; deeper local material that looked explicit-tagged, secret-bearing, internal-only, private, or too skeletal remains omitted.

Survey scope and method

  • Survey root: /Users/ericfode/src
  • Survey date: 2026-05-02
  • Coverage: 27 top-level directories under src/
  • Execution shape: exactly 10 top-level Hermes survey lanes. The runtime rejected a single 10-task batch because max_concurrent_children was 6, so the lanes ran as two batches of five while preserving exactly ten top-level lanes.
  • Recursive shape: every lane reported that deeper delegate_task was available and used. Lanes split into purpose/docs, live-work evidence, and safety/public-summary eligibility; their children reported further three-way per-directory recursion where available. The summaries were then synthesized here.
  • Evidence allowed: local README/docs, package/project files, git status, branch names, recent commits, modified/untracked file names, mtimes, plans, and visible artifacts.
  • Evidence excluded: secrets, private deployment details, hidden personal/runtime prompt details, raw explicit/provocative content, and insufficiently evidenced skeleton directories.

Ten survey lanes

  1. another-harness, are-the-astrological-signs-rascist, cardgame1
  2. deer-flow, FACEMUSIC, gas-city-but-its-just-codex
  3. gemma4-tinygrad-opt, hoid, is-codex-better
  4. is-it-formal, justfooln, kettlebellsim
  5. kimi-tests, langfuse, local-hermes
  6. meta-hermes, nnpl-external-latent-bus, nnpl-shared-bus
  7. nnpl-typed-boundary-ir, overengineeredlife, silly-pi-stuff
  8. src, testing-rl, testing-rl-hermes
  9. tinygrad, tinygrad-gemma-gemini
  10. tinygrad-gemma-kimi

Public-safe lead candidates

testing-rl

  • Evidence: README.md, pyproject.toml, docs/environment-contract.md, docs/software-risk-model.md, docs/counterfactual-testing-history.md, docs/counterfactual-verifier.md, docs/non-cheating-test-writer.md, docs/lean-verification.md, docs/alternative-test-history.md, testing_rl/*.py, tests/, benchmarks/python/*.json, formal/lakefile.lean, formal/TestingRL.lean, formal/TestingRL/Plan.lean.
  • Git: not a git worktree, so branch/commit/status evidence was unavailable.
  • Live signal: 2026-05-01 mtimes on Lean/formal docs, README, alternative-test-history docs, CLI/events code, and tests.
  • Public summary basis: a Python/Lean RL environment for training agents to write tests that expose defects, with risk modeling, counterfactual history, and formal structural checks.

FACEMUSIC

  • Evidence: docs/browser-instrument-architecture.md, ios/FACEMUSICNative/README.md, web/package.json, ml/README.md, ml/pyproject.toml, ml/configs/expression_forecast_v0_500ms.json, Cargo.toml.
  • Git: branch main, HEAD f6cf6cf; recent commits include native camera/platform hardening, iOS native camera control source, and browser face-control/stage UI stabilization.
  • Live signal: dirty worktree with about 30 entries across browser docs, iOS native files, web face-control/music-engine files, styling, and a new ML training stack; latest status mtime around 2026-04-20.
  • Public summary basis: face-controlled music instrument joining browser, iOS, audio, visual, and forecasting work. Session/capture identifiers omitted.

cardgame1 / Dungeon Steward

  • Evidence: project.godot, design/gdd/game-concept.md, prototypes/core-loop/README.md, docs/WORKFLOW-BALANCE-SIM.md, docs/engine-reference/godot/VERSION.md, LICENSE.
  • Git: branch hermes/combat-stage-art-fallback-upstream, status clean; recent commit a9a8ef6 hardens combat-stage art presentation and asset fallbacks; adjacent commits cover map viewer, floor-one layout, and hover behavior.
  • Live signal: recent balance artifact / cache mtimes after the latest commit.
  • Public summary basis: Godot 4.6-era browser-first roguelite deckbuilder prototype emphasizing deterministic combat presentation, deck/run legibility, and balance tooling.

gas-city-but-its-just-codex

  • Evidence: README.md, Cargo.toml, crate manifests, apps/operator-ui-macos/Package.swift, docs/project-formalization.md, docs/requirements/control-plane-requirements.md, docs/architecture/current-control-plane-architecture.md, docs/implementation/implementation-map.md, docs/evidence/correctness-evidence.md, formal/README.md, formal/lakefile.toml.
  • Git: branch codex/native-codex-ui, HEAD 198aefc; recent commits include Harbor task-level transfer reporting, native sandbox relaunch/operator wiring, and UI showcase workflow templates.
  • Live signal: very dirty worktree, concentrated in docs, state/config, scripts, schemas, templates, and limited Rust source areas; local/session/runtime details omitted.
  • Public summary basis: Rust Codex-native orchestration/control-plane workspace with graph ledger, operator surfaces, templates, benchmarks, and formal evidence artifacts. Operational runbook/state details withheld.

NNPL research cluster

  • nnpl-external-latent-bus: evidence from README.md, PROJECT_BRIEF.md, pyproject.toml, architecture/space/hypothesis/benchmark docs, source, tests, and report artifacts; no git metadata. It tests an external latent bus versus matched baselines.
  • nnpl-shared-bus: evidence from README.md, PROJECT_BRIEF.md, architecture/hypothesis/benchmark docs, run summaries, configs, source, and trace artifacts; no git metadata. It records a negative shared-bus result against preregistered gates.
  • nnpl-typed-boundary-ir: evidence from README.md, PROJECT_BRIEF.md, pyproject.toml, architecture/IR/hypothesis docs, source/tests, results/bakeoff.json, and tinygrad comparison results; no git metadata. It focuses on typed boundary artifacts, validation, rendering, and auditability.

Research bench / side-room candidates

  • another-harness: Lean-backed agent-harness R&D prototype with no valid HEAD commit and an untracked initial tree. Public-safe at high level as work-object/evaluator/resume/formal-semantics exploration; local MCP/config/run details omitted.
  • deer-flow: public LangGraph/Next.js/Python super-agent harness checkout with dirty local deployment/config artifacts. Public docs/source are usable, but local config is not.
  • gemma4-tinygrad-opt: local tinygrad Gemma optimization workbench with benchmark/Metal/generation scripts, package exports, and current-night artifacts. No git/root README; local logs/prompts/deployment details omitted.
  • is-codex-better: unborn Git workbench for Codex-native harness plugins: repo loops, specialist fanout, Honcho memory, transcript recall, checkpoints, jobs, and procedure promotion. Public-safe if framed as uncommitted/partially proven.
  • is-it-formal: Lean 4 + Python scaffold for grading claim formality, with JSON examples and deterministic CLI. No commits yet; public-safe as a prototype.
  • justfooln: research-and-benchmark workspace for long-horizon/tool-heavy agent harness work. Not a git repo; artifact-heavy and suitable only as a bench note.
  • local-hermes: local llama.cpp / GGUF serving setup for Hermes 4.3 36B. Public-safe only generically; no model artifact details needed.
  • src: game-development workflow/skill bundle under hidden agent-skill paths. Public-safe only as high-level process tooling.
  • silly-pi-stuff: private-marked Pi-extension sandbox plus octonion-surface browser experiment. Public-safe only as a side-room mention; hidden local prompt/config material omitted.
  • tinygrad-gemma-kimi: Git repo on opt/attention for Gemma/tinygrad optimization racing. Dirty tree with patches, reject/backups, caches, local deployment hints, and local result artifacts; summarize only as an unpolished side-room sandbox.

Held back from the public page

The following directories were surveyed but not described in project-specific public detail:

  • are-the-astrological-signs-rascist — sensitive identity/reputational research material; only generic omission is appropriate.
  • hoid — mixed creative-worldbuilding material with explicit-tagged/local/internal risk; needs curated subset before public summary.
  • kettlebellsim — clear simulation purpose, but local docs/artifacts include internal operational and credential-management references; hold back until scrubbed.
  • kimi-tests — skeletal/non-project evidence.
  • langfuse — credential/private deployment configuration signals.
  • meta-hermes — top-level scratch/meta wrapper; nested public-ish material is not the assigned root.
  • overengineeredlife — empty/no substantive evidence.
  • testing-rl-hermes — internal/Hermes-specific supervisor/evaluator and generated/local material; hold back pending separation.
  • tinygrad — empty/non-git directory.
  • tinygrad-gemma-gemini — empty/non-git directory.

Editorial synthesis

The safe publishable movement tonight clusters around three themes:

  1. testing and verification environments for agents (testing-rl, is-it-formal, parts of another-harness);
  2. embodied or game-facing interfaces (FACEMUSIC, cardgame1, selected side-room experiments);
  3. honest research benches where failures and boundaries are explicit (NNPL cluster, Gemma/tinygrad workbenches, Codex-native orchestration work).

The public page should therefore avoid pretending the tree is a single product line. It is a night desk: a testing lab, an instrument shop, a card table, and several research benches, all sharing the same insistence that claims should survive contact with artifacts.

Graph View

Backlinks